Unbiased clustering of acute-on-chronic liver failure patients using machine learning in a real-world ICU cohort

Acute-on-chronic liver failure is a complex condition with varied definitions, complicating risk stratification and targeted management. We apply unsupervised clustering to data from 1,256 patients with acute-on-chronic liver failure as defined by the North American Association for the Study of End-Stage Liver Disease. We systematically evaluate multiple algorithmic models to enable unbiased cohort stratification and to determine key clustering factors and clinical impacts across clusters. The optimal number of clusters is determined using the Partitioning Around Medoids with Ambiguous Clustering algorithm. The clusters from the best-performing nonnegative matrix factorization algorithm are selected, with the Lee’s algorithm demonstrating the best performance. Two distinct clusters are identified, showing markedly different 30-day mortality rates (70.35% vs 26.06%). Importantly, acid-base balance-related variables, including bicarbonate, pH, base excess, lactate, and anion gap, are among the primary clustering drivers. An external validation cohort, a decompensated cirrhosis cohort, and a European Association for the Study of the Liver-Chronic Liver Failure Consortium defined Acute-on-chronic liver failure patients cohort confirm consistent distributions of key clustering variables and divergence in 30-day mortality rates, supporting the role of acid-base balance variables. In summary, we show that the unbiased clustering approach successfully identifies distinct acute-on-chronic liver failure clusters with different mortality risks and emphasizes the critical role of metabolic regulation in acute-on-chronic liver failure outcomes, as well as consistent validation.